HR: 09:00h
AN: S31D-05 INVITED    [Abstracts]
TI: Optimization of smoothed-seismicity earthquake forecasts using the area skill score
AU: * Zechar, J D
EM: zechar@usc.edu
AF: University of Southern California, Department of Earth Sciences, 3651 Trousdale Pkwy, Los Angeles, CA 90089, United States
AU: Jordan, T H
EM: tjordan@usc.edu
AF: University of Southern California, Department of Earth Sciences, 3651 Trousdale Pkwy, Los Angeles, CA 90089, United States
AB: Rigorous evaluation of 5-year earthquake forecasts has begun at Collaboratory for the Study of Earthquake Predictability (CSEP) testing centers. For the California natural laboratory, initial forecasts are stated in terms of expected seismicity rate in latitude/longitude/magnitude bins. Three tests based on likelihood criteria are used to evaluate the forecasts with respect to observations of M≥5 earthquakes. These forecasts are of a general form that provides a relative ordering of estimated hazard, i.e., alarm-based forecasts. Some forecasts of the more general form cannot be tested and compared under the CSEP likelihood framework. We have developed a testing procedure based on what we call the "area skill score" of the Molchan diagram that can accommodate forecasts derived from alarm functions, including the CSEP forecasts. We understand the area skill score distribution and we use it for hypothesis testing of alarm-based forecasts. In this work, we discuss the use of the area skill score for forecast optimization. Many earthquake occurrence models, and most seismic hazard analyses, employ some form of epicentral smoothing to estimate the spatial distribution of earthquakes. In most cases, an implicit assumption is that the spatial distribution of past small earthquakes can be used to forecast the locations of future large earthquakes. We demonstrate the use of the area skill score to optimize a variety of so-called smoothed seismicity forecasts, varying the form of the smoothing kernel and the minimum magnitude of target earthquakes. We consider a series of retrospective predictability experiments in California and compare our optimization results with published results based on maximum likelihood. We also show attempts to invert for smoothing parameters using synthetic catalogs where the smoothing kernel is explicitly specified. Based on these experiments, we have constructed a set of time-invariant 5-year forecasts for California and submitted them to CSEP for testing.
UR: http://www.cseptesting.org
DE: 3245 Probabilistic forecasting (3238)
DE: 7223 Earthquake interaction, forecasting, and prediction (1217, 1242)
SC: Seismology [S]
MN: 2007 Fall Meeting